System log files are essential for recording the operational state of software and systems. Although various log anomaly detection methods exist, their performance significantly degrades when initial labeled log data is limited. In this paper, we propose LogUPS, a semi-supervised log anomaly detection method based on uncertainty-aware self-training. By leveraging a large volume of unlabeled log data through self-training, LogUPS considers both the uncertainty and confidence of pseudo-label samples during training, effectively filtering noisy pseudo-labels. This approach reduces calibration errors in the log anomaly detection model and mitigates the issue of self-training susceptibility to pseudo-label noise when initial labeled data is insufficient. Experimental results demonstrate that LogUPS outperforms existing semi-supervised log anomaly detection methods, including those based on traditional self-training.